ECG signal classification with binarized convolutional neural network.

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Cardiology

Background:

• Arrhythmias, such as atrial fibrillation (AF), are common heart rhythm disorders requiring timely detection for high-risk patients.
• Early detection of arrhythmias is critical to prevent the development of serious cardiac syndromes.

Purpose of the Study:

• To develop and evaluate deep convolutional neural network (CNN)-based algorithms for real-time arrhythmia detection.
• To investigate the feasibility of model binarization for reducing computational resource requirements while maintaining performance.

Main Methods:

• A full-precision deep convolutional neural network model was constructed and evaluated on the PhysioNet/CinC AF Classification Challenge 2017 dataset.
• Knowledge distillation was employed, using the full-precision model as a teacher to regularize the training of a binarized CNN model.
• The performance of the binarized model was compared to the full-precision model, focusing on metrics like the F1 score.

Main Results:

• The full-precision CNN model achieved state-of-the-art performance on the AF classification task.
• Binarization of the CNN model resulted in a minimal decrease in performance (F1 score from 0.88 to 0.87 on the validation set).
• The binarized model demonstrated significantly reduced computing power and memory space requirements compared to the full-precision model.

Conclusions:

• Deep convolutional neural networks are effective for real-time arrhythmia detection.
• Network binarization, regularized by knowledge distillation, offers a viable method to create computationally efficient models.
• Binarized CNNs are suitable for deployment on resource-constrained, long-term cardiac condition monitoring devices.